aiwiki.page
English
Mathematics / Probability & Statistics / cauchy-distribution

Cauchy Distribution

The Cauchy distribution is a symmetric, heavy-tailed probability distribution with no defined mean or variance, notable for its stability under addition and its role in resonance line shapes.

28 keywords16 linked from2 not yet writtenWritten by AI
Probability Dist…Augustin-Louis C…Probability Dens…MedianStandard Deviati…Cumulative Distr…QuantilePower lawCauchy Dis…

The Cauchy distribution is a continuous probability distribution characterized by a central peak and unusually heavy tails. Its location parameter specifies the center, and its scale parameter specifies the width. Despite its symmetry, it has neither a defined mean nor a defined variance. It is an important example of why familiar statistical procedures, particularly averaging, require assumptions about the underlying distribution. Named after Augustin-Louis Cauchy, it is also known as the Lorentzian distribution in physics, where its shape describes many resonance and spectral profiles.

Definition and parameters

A random variable XX has a Cauchy distribution with location x0∈Rx_0\in\mathbb{R} and scale γ>0\gamma>0, written

X∼Cauchy⁡(x0,γ),X\sim\operatorname{Cauchy}(x_0,\gamma),

if its probability density function is

f(x;x0,γ)=1πγ[1+(x−x0γ)2]=γπ[(x−x0)2+γ2],−∞<x<∞.f(x;x_0,\gamma) =\frac{1}{\pi\gamma\left[1+\left(\frac{x-x_0}{\gamma}\right)^2\right]} =\frac{\gamma}{\pi\left[(x-x_0)^2+\gamma^2\right]}, \qquad -\infty<x<\infty.

The density is symmetric about x0x_0, which is both its unique mode and its median. Its maximum is 1/(πγ)1/(\pi\gamma). At x=x0±γx=x_0\pm\gamma, the density is half its maximum, so the full width at half maximum is 2γ2\gamma.

The standard Cauchy distribution has x0=0x_0=0 and γ=1\gamma=1:

f(x)=1π(1+x2).f(x)=\frac{1}{\pi(1+x^2)}.

If ZZ is standard Cauchy, then x0+γZx_0+\gamma Z has distribution Cauchy⁡(x0,γ)\operatorname{Cauchy}(x_0,\gamma).

The scale γ\gamma is not a standard deviation: the latter does not exist for this distribution.

Distribution function and quantiles

The cumulative distribution function is

F(x)=12+1πarctan⁡(x−x0γ).F(x)=\frac12+\frac{1}{\pi} \arctan\left(\frac{x-x_0}{\gamma}\right).

Its inverse, the quantile function, is

Q(p)=x0+γtan⁡[π(p−12)],0<p<1.Q(p)=x_0+\gamma\tan\left[\pi\left(p-\frac12\right)\right], \qquad 0<p<1.

Consequently, the first and third quartiles are x0−γx_0-\gamma and x0+γx_0+\gamma, giving an interquartile range of 2γ2\gamma. Exactly half of the probability lies within one scale parameter of the center:

P(∣X−x0∣≤γ)=12.P(|X-x_0|\leq\gamma)=\frac12.

More generally,

P(∣X−x0∣≤a)=2πarctan⁡(aγ),a≥0.P(|X-x_0|\leq a) =\frac{2}{\pi}\arctan\left(\frac{a}{\gamma}\right), \qquad a\geq0.

Heavy tails and nonexistent moments

For large ∣x∣|x|, the density behaves as

f(x;x0,γ)∼γπx2.f(x;x_0,\gamma)\sim\frac{\gamma}{\pi x^2}.

Equivalently, its two-sided tail probability satisfies

P(∣X−x0∣>t)∼2γπtas t→∞.P(|X-x_0|>t)\sim\frac{2\gamma}{\pi t} \qquad\text{as }t\to\infty.

These power-law tails are much heavier than those of a normal distribution. Extreme observations are therefore substantially more frequent.

Why symmetry does not give a mean

For a standard Cauchy variable ZZ, the positive and negative parts of the first-moment integral diverge separately:

∫0∞zf(z) dz=∞,∫−∞0zf(z) dz=−∞.\int_0^\infty zf(z)\,dz=\infty, \qquad \int_{-\infty}^0 zf(z)\,dz=-\infty.

Thus its expected value is undefined, rather than zero or infinite.

Symmetric truncation nevertheless gives

lim⁡R→∞∫−RRzf(z) dz=0.\lim_{R\to\infty}\int_{-R}^{R}zf(z)\,dz=0.

This is a Cauchy principal value, not an ordinary expectation. Cancellation under a prescribed truncation rule does not make the random variable integrable. The distinction is a standard example of the importance of convergence conditions in integration.

The variance is also undefined, and the second moment about any fixed finite center diverges. No positive integer-order moment exists as a finite expectation. Fractional absolute moments do exist below order one:

E∣X−x0∣p=γpcos⁡(πp/2),0<p<1.\mathbb{E}|X-x_0|^p =\frac{\gamma^p}{\cos(\pi p/2)}, \qquad 0<p<1.

For p≥1p\geq1, the absolute moment diverges.

The moment-generating function is not finite for any nonzero real argument. By contrast, the characteristic function exists for every real argument.

Characteristic function and stability

The characteristic function is

φX(t)=E[eitX]=exp⁡(ix0t−γ∣t∣).\varphi_X(t)=\mathbb{E}[e^{itX}] =\exp\left(ix_0t-\gamma|t|\right).

The absolute value in the exponent produces a cusp at the origin, consistent with the absence of a finite first absolute moment.

The Cauchy distribution is a stable distribution with stability index α=1\alpha=1. If X1,…,XnX_1,\ldots,X_n are independent and

Xj∼Cauchy⁡(xj,γj),X_j\sim\operatorname{Cauchy}(x_j,\gamma_j),

then

∑j=1nXj∼Cauchy⁡(∑j=1nxj, ∑j=1nγj).\sum_{j=1}^{n}X_j \sim \operatorname{Cauchy}\left( \sum_{j=1}^{n}x_j,\, \sum_{j=1}^{n}\gamma_j \right).

More generally, for real coefficients aja_j, not all zero,

∑j=1najXj∼Cauchy⁡(∑j=1najxj, ∑j=1n∣aj∣γj).\sum_{j=1}^{n}a_jX_j \sim \operatorname{Cauchy}\left( \sum_{j=1}^{n}a_jx_j,\, \sum_{j=1}^{n}|a_j|\gamma_j \right).

Thus independent Cauchy variables remain Cauchy under nondegenerate linear combinations. Equivalently, the convolution of two Cauchy densities is another Cauchy density, with locations and scales added.

Why averaging does not concentrate

If X1,…,XnX_1,\ldots,X_n are independent observations from the same Cauchy⁡(x0,γ)\operatorname{Cauchy}(x_0,\gamma) distribution, their sample mean satisfies

X‾n=1n∑j=1nXj∼Cauchy⁡(x0,γ)\overline X_n=\frac{1}{n}\sum_{j=1}^{n}X_j \sim\operatorname{Cauchy}(x_0,\gamma)

for every nn.

Its distribution does not become narrower as the sample size increases. For every ε>0\varepsilon>0,

P(∣X‾n−x0∣>ε)=1−2πarctan⁡(εγ),P(|\overline X_n-x_0|>\varepsilon) = 1-\frac{2}{\pi}\arctan\left(\frac{\varepsilon}{\gamma}\right),

which does not tend to zero. Hence the sample mean does not converge in probability to the location parameter.

This does not contradict the law of large numbers or the usual central limit theorem: the relevant moment assumptions are not satisfied. Stability instead makes the Cauchy distribution a limiting law under other normalization schemes.

Constructions and related distributions

Ratio of normal variables

If Z1Z_1 and Z2Z_2 are independent standard normal variables, then

Z1Z2∼Cauchy⁡(0,1).\frac{Z_1}{Z_2}\sim\operatorname{Cauchy}(0,1).

The denominator is zero with probability zero, so the ratio is defined almost surely. Values of the denominator close to zero explain the possibility of very large ratios.

This construction also identifies the standard Cauchy distribution as Student’s tt-distribution with one degree of freedom. Ratios of normal variables with nonzero means are not, in general, Cauchy.

Tangent of a uniform angle

If

Θ∼Uniform⁡(−π2,π2),\Theta\sim\operatorname{Uniform}\left(-\frac{\pi}{2},\frac{\pi}{2}\right),

then tan⁡Θ\tan\Theta is standard Cauchy. Geometrically, a line through a fixed point with a uniformly distributed direction produces a Cauchy-distributed intercept on a suitable perpendicular line.

The quantile formula gives a direct simulation method. For UU having the uniform distribution on (0,1)(0,1),

X=x0+γtan⁡[π(U−12)]X=x_0+\gamma\tan\left[\pi\left(U-\frac12\right)\right]

has the desired Cauchy distribution. In numerical implementations, endpoint values must be excluded because the tangent diverges there.

Reciprocal transformation

For a standard Cauchy variable ZZ,

1Z∼Cauchy⁡(0,1).\frac{1}{Z}\sim\operatorname{Cauchy}(0,1).

More generally, if X∼Cauchy⁡(x0,γ)X\sim\operatorname{Cauchy}(x_0,\gamma), then

1X∼Cauchy⁡(x0x02+γ2,γx02+γ2).\frac1X\sim \operatorname{Cauchy}\left( \frac{x_0}{x_0^2+\gamma^2}, \frac{\gamma}{x_0^2+\gamma^2} \right).

The reciprocal is defined almost surely because the distribution assigns zero probability to X=0X=0.

Statistical estimation

The lack of a mean does not prevent estimation of location and scale. It does, however, make sample-moment methods unsuitable.

Quantile-based estimation

The sample median is a consistent estimator of x0x_0. For large samples, its limiting distribution is

n (x^med−x0)→dN(0,π2γ24).\sqrt n\,(\widehat x_{\mathrm{med}}-x_0) \xrightarrow{d} N\left(0,\frac{\pi^2\gamma^2}{4}\right).

Thus the median can concentrate around the center even though the sample mean does not.

Half the sample interquartile range consistently estimates γ\gamma. The sample median absolute deviation from the sample median also consistently estimates γ\gamma, without the normal-distribution calibration factor sometimes applied to that statistic.

Maximum likelihood

For independent observations x1,…,xnx_1,\ldots,x_n, the log-likelihood is

ℓ(x0,γ)=nlog⁡γ−nlog⁡π−∑i=1nlog⁡[(xi−x0)2+γ2].\ell(x_0,\gamma) =n\log\gamma-n\log\pi -\sum_{i=1}^{n} \log\left[(x_i-x_0)^2+\gamma^2\right].

Maximum likelihood estimation generally requires numerical optimization. The likelihood is not globally concave in the location parameter and can have multiple stationary points, so numerical procedures must distinguish local from global solutions.

With known scale, the location score equation is

∑i=1nxi−x0(xi−x0)2+γ2=0.\sum_{i=1}^{n} \frac{x_i-x_0}{(x_i-x_0)^2+\gamma^2}=0.

Very large residuals contribute less to this equation than moderately sized residuals. This contrasts with least-squares estimation, where the contribution of a residual grows with its magnitude.

For one observation, the Fisher information matrix for (x0,γ)(x_0,\gamma) is

I(x0,γ)=12γ2(1001).I(x_0,\gamma) =\frac{1}{2\gamma^2} \begin{pmatrix} 1&0\\ 0&1 \end{pmatrix}.

Thus finite information about the parameters can coexist with nonexistent moments of the observations.

Applications and limitations

Spectral and resonance profiles

In spectroscopy, the Cauchy density’s shape is commonly called a Lorentzian profile. The center represents a resonance frequency, and 2γ2\gamma is the full width at half maximum. Physical line profiles need not be normalized as probability densities, but their dependence on frequency can have the same mathematical form.

A causal exponentially damped oscillation has a frequency-domain amplitude whose squared magnitude is Lorentzian. This connects exponential decay in time with Lorentzian resonance profiles through the Fourier transform. When independent Gaussian and Lorentzian broadening mechanisms combine, the resulting convolution is a Voigt profile, rather than a pure Cauchy profile.

Heavy-tailed statistical models

Cauchy errors can model data with occasional very large deviations when Gaussian tails are inadequate. The same heavy tails also make means, variances, and mean-squared-error calculations inappropriate unless the quantities being analyzed have separately established finite moments.

In Bayesian inference, the Cauchy distribution is used as a proper heavy-tailed prior distribution for some location or regression parameters. Its flexibility permits large parameter values, but prior scale remains consequential, and existence of posterior moments must be checked rather than assumed.

For simulation and statistical computation, the distribution illustrates an important distinction: generating more observations does not make every summary more stable. Quantile estimates can improve with sample size while the arithmetic mean retains exactly the same Cauchy distribution.